-
Notifications
You must be signed in to change notification settings - Fork 6
/
main_GAN.py
190 lines (144 loc) · 6.51 KB
/
main_GAN.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
import os
from option import args
os.environ['CUDA_VISIBLE_DEVICES'] = args.cuda_name
import torch
import torch.optim as optim
import os
import torch.nn as nn
from loss import gan
from data import dataset
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
import os
from option import args
import torchvision
import models
from utils.metric import torch_psnr
from models.discriminator import LightCNN_9Layers, VGGFeatureExtractor, LightCNN_9Layers_fft
device = torch.device(args.device)
def to_device(sample, device):
for key, value in sample.items():
if key != 'img_name':
sample[key] = value.to(device, non_blocking=True)
return sample
epochs = args.epochs
start_epoch = 0
lr = args.lr
model = models.get_model(args)
D = LightCNN_9Layers()
Vgg = VGGFeatureExtractor()
fftD = LightCNN_9Layers_fft()
fftD = fftD.to(device, non_blocking=True)
D = D.to(device, non_blocking=True)
Vgg = Vgg.to(device, non_blocking=True)
Vgg.eval()
pre_trained = torch.load(args.load)
model.load_state_dict(pre_trained)
model = model.to(device, non_blocking=True)
writer = SummaryWriter('./logs/{}'.format(args.writer_name))
traindata = dataset.Data(root=os.path.join(args.data_path,"CelebA/train"), args=args, train=True)
valdata = dataset.Data(root=os.path.join(args.data_path,'CelebA/val'), args=args, train=False)
testdata1 = dataset.Data(root=os.path.join(args.data_path,'CelebA/test'), args=args, train=False)
testdata2 = dataset.Data(root=os.path.join(args.data_path,'helen/test'), args=args, train=False)
trainset = DataLoader(traindata, batch_size=args.batch_size, shuffle=True, num_workers=32)
valset = DataLoader(valdata, batch_size=args.batch_size, shuffle=False, num_workers=1)
testset1 = DataLoader(testdata1, batch_size=args.batch_size, shuffle=False, num_workers=1)
testset2 = DataLoader(testdata2, batch_size=args.batch_size, shuffle=False, num_workers=1)
class AMPLoss(nn.Module):
def __init__(self):
super(AMPLoss, self).__init__()
self.cri = nn.L1Loss()
def forward(self, x, y):
x = torch.fft.rfft2(x, norm='backward')
x_mag = torch.abs(x)
y = torch.fft.rfft2(y, norm='backward')
y_mag = torch.abs(y)
return self.cri(x_mag,y_mag)
class PhaLoss(nn.Module):
def __init__(self):
super(PhaLoss, self).__init__()
self.cri = nn.L1Loss()
def forward(self, x, y):
x = torch.fft.rfft2(x, norm='backward')
x_mag = torch.angle(x)
y = torch.fft.rfft2(y, norm='backward')
y_mag = torch.angle(y)
return self.cri(x_mag, y_mag)
def eval_model(model, dataset, name, epoch, args):
model.eval()
val_psnr_dic = 0
val_ssim_dic = 0
os.makedirs(os.path.join(args.save_path, args.writer_name, 'result'), exist_ok=True)
for batch, data in enumerate(dataset):
sr = model(to_device(data, device))
psnr_c, ssim_c = torch_psnr(data['img_gt'], sr['img_out'])
val_psnr_dic = val_psnr_dic + psnr_c
val_ssim_dic = val_ssim_dic + ssim_c
print("Epoch：{}, {}, psnr: {:.3f}".format(epoch+1, name, val_psnr_dic/(len(dataset))))
writer.add_scalar("{}_psnr_DIC".format(name), val_psnr_dic/len(dataset), epoch)
writer.add_scalar("{}_ssim_DIC".format(name), val_ssim_dic / len(dataset), epoch)
def train_model(model, trainset, epoch, args):
model.train()
criterion1 = nn.L1Loss().to(device, non_blocking=True)
amploss = AMPLoss().to(device, non_blocking=True)
phaloss = PhaLoss().to(device, non_blocking=True)
Gan_loss = gan.GANLoss().to(device, non_blocking=True)
optimizer = optim.Adam(params=model.parameters(), lr=lr, betas=(0.9, 0.99), eps=1e-8)
optimizer_D = optim.Adam(params=D.parameters(), lr=lr, betas=(0.9, 0.99), eps=1e-8)
optimizer_D_fft = optim.Adam(params=fftD.parameters(), lr=lr, betas=(0.9, 0.99), eps=1e-8)
train_loss = 0
d_loss = 0
d_loss_fft = 0
for batch, data in enumerate(trainset):
sr = model(to_device(data, device))
l1_loss = criterion1(sr['img_out'], data['img_gt']) + \
args.fft_weight * amploss(sr['img_fre'], data['img_gt']) + args.fft_weight * phaloss(
sr['img_fre'],
data[
'img_gt']) + \
criterion1(sr['img_fre'], data['img_gt'])
SR_feature = Vgg(sr['img_out']).detach()
HR_feature = Vgg(data['img_gt']).detach()
perceptual_loss = criterion1(HR_feature, SR_feature)
SR_pred = D(sr['img_out'])
# HR_pred = D(data['img_gt']).detach()
loss_g_GAN = Gan_loss(SR_pred, True)
SR_pred_fft = fftD(sr['img_out'])
loss_g_GAN_fft = Gan_loss(SR_pred_fft, True)
loss_g = 0.0005 * loss_g_GAN + l1_loss + 0.1 * perceptual_loss + args.fftd_weight * loss_g_GAN_fft
train_loss = train_loss + loss_g.item()
optimizer.zero_grad()
loss_g.backward()
optimizer.step()
optimizer_D.zero_grad()
optimizer_D_fft.zero_grad()
for p in D.parameters():
p.requires_grad = True
for p in fftD.parameters():
p.requires_grad = True
HR_pred = D(data['img_gt'])
SR_pred = D(sr['img_out'].detach())
loss_d = Gan_loss(HR_pred, True) + Gan_loss(SR_pred, False)
d_loss += loss_d.item()
loss_d.backward()
optimizer_D.step()
SR_pred_fft = fftD(sr['img_out']).detach()
HR_pred_fft = fftD(data['img_gt'])
loss_d_fft = args.fftd_weight * Gan_loss(HR_pred_fft, True) + args.fftd_weight * Gan_loss(SR_pred_fft, False)
d_loss_fft += loss_d_fft.item()
loss_d_fft.backward()
optimizer_D_fft.step()
print("Epoch：{} loss: {:.3f}".format(epoch + 1, train_loss / (len(trainset)) * 255))
writer.add_scalar('train_loss', train_loss / (len(trainset)) * 255, epoch + 1)
writer.add_scalar('d_loss', d_loss / (len(trainset)) * 255, epoch + 1)
writer.add_scalar('d_loss_fft', d_loss_fft / (len(trainset)) * 255, epoch + 1)
os.makedirs(os.path.join(args.save_path, args.writer_name), exist_ok=True)
os.makedirs(os.path.join(args.save_path, args.writer_name, 'model'), exist_ok=True)
torch.save(model.state_dict(),
os.path.join(args.save_path, args.writer_name, 'model', 'epoch{}.pth'.format(epoch + 1)))
if __name__ == "__main__":
for i in range(epochs):
train_model(model, trainset, i, args)
eval_model(model, valset, "val", i, args)
eval_model(model, testset1, "CelebA", i, args)
eval_model(model, testset2, "Helen", i, args)